Multiple Instance Filtering
Kamil Wnuk, Stefano Soatto · 2011
We propose a robust filtering approach based on semi-supervised and mul-tiple instance learning (MIL). We assume that the posterior density would be unimodal if not for the effect of outliers that we do not wish to explic-itly model. Therefore, we seek for a point estimate at the outset, rather than a generic approximation of the entire posterior. Our approach can be thought of as a combination of standard finite-dimensional filtering (Ex-tended Kalman Filter, or Unscented Filter) with multiple instance learning, whereby the initial condition comes with a putative set of inlier measure-ments. We show how both the state (regression) and the inlier set (classi-fication) can be estimated iteratively and causally by processing only the current measurement. We illustrate our approach on visual tracking prob-lems whereby the object of interest (target) moves and evolves as a result of occlusions and deformations, and partial knowledge of the target is given in the form of a bounding box (training set). 1